Available now. Senior AI roles.

AI engineer.Production systems.

I design, build and run production LLM systems. RAG pipelines, agentic workflows, evaluation harnesses. Currently CTO at Atlas Wealth.

  • 4h → 10minplan generation
  • 2×new-client rate
  • 4k+documents indexed
  • 20+model versions, zero regressions
Selected work Contact
Portrait of Nelson Dell
Nelson DellPorto Alegre, Brazil
Currently
CTO, Atlas Wealth
Experience
Engineering since 2020
Focus
RAG, evals, MLOps
Based
Porto Alegre, BR
01. AboutShort version

Engineer. I build systems that survive contact with real users.

I work in the gap between a notebook and a production system. Data contracts, evaluation harnesses, retrieval guardrails, deploy pipelines. The infrastructure that turns a prompt into a service a business can depend on.

Currently founding partner, CTO and sole engineer of Atlas Wealth, where I built and shipped Atlas Advisor: a RAG platform used daily by all six advisors that cut financial plan generation from four hours to under ten minutes. Previously led a commercial team at Faz Capital, co-founded PureAI, shipped predictive models at Bateleur, and started as a front-end engineer at Ilegra.

I care about shipped systems over benchmarks, evaluation over intuition, and reliable infrastructure over clever architecture. I read the failure traces. I write the runbooks. I am on call when the agent loops.

Based in Porto Alegre. Currently completing a post-graduate specialisation in Applied AI Software Engineering at Anhanguera.

Based
Porto Alegre, BRRio Grande do Sul.
Role
Founding Partner & CTOAtlas Wealth. Fintech.
Education
B.Sc. Data Science & AIPUCRS, 2024
Post-grad
Applied AI Software Eng.Anhanguera, 2026–27
Languages
English, Português, EspañolFluent / native / advanced
Status
Open to senior rolesFull-time. Remote-friendly.
02. Selected work2023 / 2026

Five projects. Shipping or in build.

No. 001FeaturedRAG / Production

Atlas Advisor

LLM investment advisor. Atlas Wealth, 2025.

Retrieval-augmented advisor over 4,000+ financial documents: fund prospectuses, regulatory filings, internal allocation memos. Produces personalised client plans with citation-traceable recommendations, used daily by all six advisors at Atlas Wealth.

pgvector retrieval with re-ranking, hallucination guardrails, and a continuous evaluation harness with golden datasets and regression tests, used to promote 20+ model versions to production without quality regressions.

Served via FastAPI on AWS Bedrock and Lambda. Cut financial plan generation from four hours to under ten minutes.

PythonLangChainpgvectorAWS BedrockMLflowFastAPIPostgreSQL
Live. productionatlas-advisor.prod
4h→10m
Plan
generation
20+
Model versions
promoted
4k+
Documents
indexed
No. 002FeaturedAgentic RAG / Solo

Fisko

Tax-reform copilot over LC 214/2025. Independent project, 2026.

Independent RAG system over Brazil's tax reform (LC 214/2025: IBS, CBS, IS). Ingests the law, regulatory notes and CARF rulings, and produces a diagnostic report that shows how a given business is exposed to the transition.

Ragas evaluation suite with faithfulness, answer-relevancy and context-precision gates wired into CI. Free public diagnostic, paid PDF report at R$47.

Built solo, end to end: retrieval, evals, backend, frontend, payments, hosting. MVP v1.0 live at fisko.pro since April 2026.

PythonLangChainRagaspgvectorPostgresNext.js
Livefisko.pro
LC214
Tax reform
covered
R$47
Paid report
price
v1.0
MVP live
Apr 2026
No. 003Vertical SaaS / Build

Prosa

Direct-order platform for Brazilian restaurants. useprosa.com.

Lets gourmet hamburger restaurants in Brazil escape iFood's 12 to 27% commission and own the customer relationship. Premium digital menu on the restaurant's own domain, direct ordering via WhatsApp with Pix, automated retention campaigns, and an economy dashboard that shows commission saved and customer LTV.

Inspired by Owner.com, adapted for the Brazilian reality where Instagram and WhatsApp, not Google, are the discovery channels. Pilot in Porto Alegre with around 100 target restaurants, regional TAM of 400 in South Brazil, national TAM of around 12,000.

Pricing tiers from R$197 to R$897. Target 30 restaurants in 6 months, R$60 to 100k MRR in 18.

Next.jsWhatsApp Business APIPix / Mercado PagoPostgresAI generation
MVP in builduseprosa.com
12k
Restaurants
national TAM
25%
Avg. commission
removed
R$100k
Target MRR
18 months
No. 004Vertical SaaS / Pre-launch

Mecaniko

Operating system for the independent oficina. mecaniko.com.br.

Vertical SaaS for independent Brazilian auto repair shops. Mobile-first, WhatsApp-native, AI-assisted. Built for shops with 3 to 12 bays running R$80k to R$500k monthly revenue, currently stuck on a mix of paper service orders, manual Pix, and decade-old ERPs.

WhatsApp-native customer communication, Pix with automatic confirmation, NFS-e integration, AI-assisted service order creation and diagnostics. Designed for someone in a graxa-stained shirt, on the floor, on a phone.

Visual and product peers: Shopmonkey, ServiceTitan, ToastTab. Premium positioning at R$197 to R$997 against R$50 to R$300 incumbents.

Next.jsWhatsApp Business APIPixNFS-eAI assist
Pre-launchmecaniko.com.br
3–12
Bays
per shop
R$500k
Top-of-band
monthly revenue
R$997
Top tier
monthly
No. 005Open Source / Tooling

agentic-starter

Production foundation for AI systems that don't make things up. MIT.

Opinionated open source template for shipping LLM systems with Claude Code. Codifies the patterns I kept rebuilding project to project: tight Skills with frontmatter for on-demand context, CLAUDE.md under 100 lines, deterministic hooks, isolated TDD subagents, model routing, guardrails (ConfidenceValidator, SourceValidator, ToolRoutingValidator), and an AgentTracer that detects hallucination automatically.

Spec-Driven Development methodology, evals folder split by fast and llm markers, lessons.md self-improvement loop, CI that blocks broken commits.

PythonNext.jsuvpytestClaude Code
v0.1 / MITgithub.com/Juniordell/agentic-starter
<100
Lines in
CLAUDE.md
3
Optional
modules
03. ExperienceReverse chronological

Software, data and AI engineering. Two founding teams.

2025 / Present

AI Engineer, Founding Partner & CTO

Atlas Wealth. Fintech.

Built and shipped Atlas Advisor, a RAG platform used daily by all six advisors, cutting financial plan generation from four hours to under ten minutes. Sole engineer on the product: retrieval pipeline (pgvector over 4k+ documents), AWS infrastructure, CI/CD, evaluation harness, and the TypeScript frontend.

RAG, pgvectorAWS BedrockMLflow, AirflowEval harnesses
2023 / 2024

Commercial Leader

Faz Capital

Led a commercial team of 7 advisors managing R$90M+ in AUM, exceeding revenue targets by 60% YoY. Built data-driven sales playbooks around client segmentation and conversion funnels. The manual planning workflow I saw here became the product thesis for Atlas Advisor.

Team leadershipSales strategyClient segmentation
2022 / 2023

Co-Founder & AI Engineer

PureAI

Co-founded an applied AI startup and fine-tuned GPT models for contract Q&A in late 2022, before RAG tooling went mainstream. Delivered document Q&A solutions for clients including the City of Porto Alegre and Hospital Ernesto Dornelles.

GPT fine-tuningContract Q&ANLP
2021 / 2022

Data Scientist

Bateleur

Automated weather data collection from 4 meteorological sources for an agribusiness client, cutting manual gathering time by 80%. Built 7+ data pipelines with Selenium and Python, and automated Power BI reporting that took recurring reports from three hours to ten minutes.

Selenium, Pandasscikit-learnPower BI, SQL
2020 / 2021

Software Developer

Ilegra

Built responsive React and TailwindCSS interfaces for enterprise client projects, contributing to 6 shipped features across 2 products. Optimised component rendering, cutting page load time by 15%, and wrote reusable component documentation that onboarded 5 new developers.

React, TSTailwindCSSREST APIs
04. StackIn production weekly

Tools I ship with.

AI & ML

  • LangChainDaily
  • LangGraphDaily
  • AWS BedrockAPI
  • RAG pipelinesRetrieval
  • pgvectorVector
  • LLM evaluationGolden sets
  • Fine-tuningPeriodic
  • scikit-learnClassical
  • MLflowTracking

Data & Infrastructure

  • PythonNative
  • FastAPIDaily
  • PostgreSQL, SQLDaily
  • TypeScript / JSFrontend
  • AWS Bedrock, LambdaProduction
  • DockerDaily
  • AirflowPipelines
  • Git, GitHub ActionsCI/CD

Analytics & Method

  • Pandas, NumPyDaily
  • Statistical modellingApplied
  • A/B testing, evalsApplied
  • Power BIReporting
  • Data contractsDiscipline
  • Production MLDiscipline
  • Eval harnessesDiscipline
05. ContactDirect lines

Get in touch.

Emailnelson.delljr@gmail.comLinkedIn/in/nelson-dellGitHub/JuniordellRésuméDownload PDF